Learning Model Confidence Weighting for Picking Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The recognition accuracy of models trained through teachingless learning for picking robots may decrease due to the uniform treatment of candidates as false, even if they are potentially adoptable as gripping positions, leading to low confidence outputs for viable options.
Innovation Solution
A learning method that calculates a second value by weighting the first value to decrease as confidence increases, updating the model to minimize errors for candidates with high confidences, thereby improving recognition accuracy by differentiating between true and false candidates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If teachingless learning is used to train the model autonomously, then the model can perform image recognition without manual teaching, but recognition accuracy decreases because all candidates are uniformly treated as false
Solution Approach 1:
The patent applies local quality by differentiating the treatment of candidate regions based on their confidence scores. High-confidence candidates are treated differently from low-confidence candidates in the error calculation process. Specifically, the error for high-confidence candidates is calculated differently to preserve potentially correct detections, while low-confidence candidates are treated as false positives. This localized differentiation in error treatment resolves the contradiction by maintaining automation while improving accuracy.
Solution Approach 2:
The patent changes the parameter of error calculation based on confidence levels. Instead of using a uniform error calculation for all candidates, the system adjusts the error calculation method according to the confidence score threshold. Candidates above the threshold have their errors calculated to preserve potential true positives, while candidates below the threshold are treated as false. This parameter-based differentiation resolves the accuracy issue while maintaining autonomous operation.
2Device complexity
If all candidates are uniformly treated as false in teachingless learning, then the learning process is simplified, but viable gripping position candidates are incorrectly penalized
Solution Approach 1:
The patent introduces a confidence score threshold parameter that divides candidates into two groups: high-confidence candidates (above threshold) and low-confidence candidates (below threshold). This parameter-based classification allows the system to treat different candidates differently in the error calculation, improving reliability without significantly increasing complexity. The threshold parameter provides a simple yet effective mechanism for differentiated treatment.
Solution Approach 2:
The patent segments the candidate set into two distinct groups based on confidence scores: high-confidence candidates and low-confidence candidates. This segmentation allows the system to apply different error calculation strategies to each group. High-confidence candidates are treated as potentially correct and their errors are calculated differently, while low-confidence candidates are treated as false positives. This segmentation improves reliability while maintaining relatively simple learning process complexity.
Data Source
AI summary
A learning method is performed by a computer. The method includes: inputting a first image to a model, which outputs, from an input image, candidates for a specific region and confidences indicating probabilities of the respective candidates being the specific region, to cause the model to output a plurality of candidates for the specific region and confidences for the respective candidates; calculating a first value for each of candidates whose confidences do not satisfy a certain criterion among the candidates output by the model, the first value increasing as the confidence increases; calculating a second value obtained by weighting the first value such that the second value decreases as the confidence increases; and updating the model such that the second value decreases.